As computer vision enables the robot to be aware of its human counterpart, such algorithms could help machines to achieve human-like interaction. However, many video tracking algorithms are not able to cope with some robot vision requirements. The articulated tracking system we develop solves some of those issues. It relies on model-based algorithms, which we believe are more suitable to robot vision than appearance-based ones. Indeed, as they update all the relevant parameters of a surrounding world model, results include some knowledge of the camera and objects relative positions. Our system relies on 3D model silhouette matching and runs in real time. We increase the algorithm robustness by introducing a pre-processing step based on image moments. This allows the iteration refinement to start in a better position by roughly estimating the body motion from one frame to the next.
Architecture and Performance Evaluation of a Model-Based Robot Vision VLSI Processor for 3-D Instrumentation
